Do Japanese speakers perceive nonexistent vowels innon-native consonant clusters?

S. Funatsua, S. Imaizumib, M. Fujimotoc, A. Hashizumed and K. Kurisud

aThe Prefectural Univ. of Hiroshima, 1-1-71 Ujinahigashi Minami-ku, 734-8558 Hiroshima,Japan

bThe Prefectural Univ. of Hiroshima, 1-1 Gakuen-machi, 723-0053 Mihara, JapancThe National Inst. for Japanese Language, 10-2 Midori-machi, 190-8561 Tachikawa, Japan

dHiroshima Univ., 1-2-3 Kasumi Minami-ku, 734-8553 Hiroshima, Japanfunatsu@pu-hiroshima.ac.jp

Acoustics 08 Paris

651

Vowel epenthesis is a well known phenomenon that non-native speakers insert epenthetic vowels inside non-native consonant clusters. Vowel epenthesis is assumed as perceptual "illusory vowels" (Dupoux et al. 1999). We analyzed vowel epenthesis shown by native Japanese speakers during reading and repetition tasks for non-native consonant clusters, and analyzed their brain responses using magnetoencephalographic methods. Under the reading task, in which subjects read English words and nonsense words, native Japanese speakers inserted vowel /o/ after /t/ and /d/ in consonant clusters, and vowel /u/ after other consonants. Under the repetition task, in which subjects repeated utterances of a native English speaker, native Japanese speakers did not produce epenthetic vowels, with a few exceptions. The length of exceptional epenthetic vowels found under the repetition task was shorter than those under the reading task. The magnetoencephalographic mismatch responses were elicited by epenthetic vowels, suggesting that the native Japanese detected epenthetic vowels as real segments. Based on these results, we conclude that vowel epenthesis by native Japanese speakers did not arise from "illusory vowels", but rather from the difficulty of articulating consonant clusters which are not found in Japanese, and/or Japanese phonological rules.

1 Introduction

It is well known that second language learners have difficulties in pronunciation of some phonemes which do not exist in their native languages. For instance, French nasal vowels are hard to pronounce correctly for Japanese learners. Some phonemes, even belonging to the native language, may be difficult to pronounce depending on the contexts which the phonotactics of the language do not allow the phoneme sequences. English speakers, for instance, pronounce “zg” and “zb” in Polish /z g/ and /z b/, respectively [1]. Japanese speakers often insert vowels in consonant clusters (e. g. English word “drama” may be pronounced as /dorama/), possibly because Japanese does not allow the consonant sequences due to its phonotactics. Dupoux et al. reported that Japanese speakers perceive non-existent “illusory vowel /u/” in consonant clusters based on their phoneme detection tasks and ABX tests [2, 3]. Denaene-Lambertz et al. reported that using quasi-oddball paradigm (igmo vs. igumo) in electroencephalographic experiment they could not observe mismatch like pattern in Japanese subjects, but observed it in French subjects [4]. However, in the stimuli used in Dehaene-Lambertz’s study and Dupoux’s study, coarticulation with following vowels was not taken into account. They removed the vowel segment between the consonants to create the target consonant clusters. Therefore, some coarticulation effects from the deleted vowel segment might remain in the consonants, which may generate different timbres compared to the consonants articulated without the vowel segment. Consonant sequences and closed syllables are not allowed in Japanese. Vowel epenthesis is frequently observed in loan words, and vowel /o/ is inserted after /t/ and /d/, vowel /i/ is inserted after /t / and /d /, and vowel /u/ is inserted after other consonants [5]. The epenthetic vowels are chosen based on faithfulness constraints i. e. “faithfulness constraints require the correspondence between the input and output” [6, 7]. In this study, we investigated vowel epenthesis in Japanese in two aspects, production and perception. We carried out four experiments. Reading test 1 was done to confirm the aspects of vowel epenthesis of native Japanese speakers. A repetition test examined the relationship between perception and production. Reading test 2 confirmed the robustness of vowel epenthesis. Then a

magnetoencephalographic test revealed the potential of the detection of epenthetic vowels in Japanese speakers.

2 Reading test 1

2.1 Method

The speech samples were English words which have a consonant cluster including /t/ or /d/ at the word initial position and non-words which have a consonant cluster at the word initial position. These target words and non-words written in Roman alphabet as shown in Table A were presented to 16 native Japanese female speakers to read aloud in a sound proof room. To verify the effects of words vs. non-words utterance order, the subjects were divided into two groups, A and B. Group A (8 subjects) read English words first and non-words next, Group B read in the reverse order. The utterances were recorded by a solid state recorder (Marantz PMD671) at an 11.025 kHz sampling rate. Epenthetic vowels and their lengths were measured by visual inspection of the spectrogram on a computer. Table A English word

drill drama drive dress dry strike straight stripe trouble truck trend try

Non-word

edmo etmo dra tra gra cra

bra pra

2.2 Results

Table 1 and 2 shows the result of the reading test 1. A~H and O~V indicate individual subjects. In most of the words and non-words, subjects inserted vowels in consonant clusters. However, in the case of /s/-/t/ in cluster /str/, no voiced vowel is observed. In /s/-/t/, vowel devoicing might occur [8]. Vowel /o/ was inserted after /t/ and /d/, whereas vowel /u/ was inserted after other stop consonants (/p/, /b/, /k/, /g/). There are not any significant differences between the results shown in Table 1 with Table 2, suggesting no significant effects of the utterance order, words vs. non-words. Table 3 shows the mean values of the pooled data of Table 1 and 2. Vowel /o/ length inserted in the consonant

Acoustics 08 Paris

652

cluster /tr/ was significantly shorter than that in the consonant cluster /dr/ (p<0.0001) and the /str/ (p<0.0001). Vowel length in the /dr/ did not significantly differ from that in the /str/ (p=0.6046). Figure 1 shows spectrograms of “straight”. The spectrogram (a) is an utterance of a native speaker of American English and (b) is of a native Japanese subject E. As shown in Fig. 1 (b), no voiced vowel is observed between /s/ and /t/, while a vowel /o/ between /t/ and /r/ is observed. Figure 2 shows non-word “dra” and “gra”. Vowel /o/ is observed between /d/ and /r/, whereas /u/ is inserted between /g/ and /r/. Namely, it is confirmed that vowel /o/ inserted after /t/ and /d/ and vowel /u/ inserted after the other stop consonants as many previous studies demonstrated.

A B C D E F G H meandrill 107 84 81 63 62 64 86 107 79.9

drama 52 55 53 48 55 21 52 56 52.2drive 44 [u] 31 66 42 52 50 63 51 53.4dress 66 74 57 59 39 44 49 82 58.6dry 47 63 60 59 73 47 41 80 58.8

strike 36 40 58 45 60 63 68 60 57.2straight 65 42 52 58 76 55 60 54 59.7stripe 66 50 86 58 70 44 70 46 61.3

trouble 37 0 36 35 24 21 28 39 28.1truck 29 32 49 40 24 21 39 56 35.3trend 23 37 42 59 37 23 44 60 38.7try 42 0 53 38 55 31 31 51 38.3

edmo 121 37 69 67 72 54 91 [u] 88 72.9etmo 58 39 69 51 54 67 62 [u] 92 54.5dra 89 78 63 83 55 99 62 73 76.9tra 45 60 82 56 42 55 52 85 59.3gra 79 74 57 58 41 140 49 80 71.2cra 53 57 24 29 24 57 42 65 45.6bra 73 105 47 63 73 65 47 63 71.7pra 45 33 60 45 39 52 [a] 68 85 51.3

Table 1 Epenthetic vowel lengths in the reading test 1 (Group A). (ms) Epenthetic vowels were inserted in underlined consonants. A-H indicate individual subjects. Subjects usually inserted /o/ in “drill”-“tra”, while they inserted vowel /u/ in “gra”-“pra”.

O P Q R S T U V meandrill 59 54 59 81 100 [u] 92 71 71 70.7

drama 92 55 36 60 74 [u] 37 [#] 56 39 59.3drive 22 31 62 28 48 [u] 43 43 [u] 17 39.0dress [u] 31 38 69 62 64 [u] 54 [u] 27 51 56.8dry [u] 29 47 45 0 57 [u] 66 0 36 30.8

strike 36 48 33 83 66 [u] 44 53 37 50.9straight 19 74 55 x 41 [u] 49 31 46 44.3stripe 10 41 45 x 45 90 32 44 36.2

trouble 19 29 50 24 22 56 [#] 39 0 28.6truck 16 43 54 22 33 [u] 34 46 29 34.7trend 0 24 40 48 40 32 [u] 17 32 30.9try 28 36 38 0 50 [u] 54 0 27 25.6

edmo 62 73 69 52 x [u] 53 0 43 49.8etmo 66 57 31 59 92 [u] 56 0 39 49.1dra [u] 21 88 52 88 [u] 129 [u] 72 34 49 62.2tra 43 62 48 64 95 [u] 54 39 43 56.3gra 35 120 55 83 158 70 20 26 70.9cra 38 112 55 130 [a] 119 43 24 29 61.6bra 60 71 57 62 119 77 66 27 67.4pra 22 119 24 40 95 49 34 31 51.8

Table 2 Epenthetic vowel lengths in the reading test 1 (Group B). (ms) O-V indicate individual subjects. X indicates pronunciation error. [#] = [ ]

Figure 1 Spectrograms of “straight”.

(a) a male speaker of native English, (b) subject E.

Figure 2 Spectrograms of “dra” and “gra” in subject D.

3 Repetition test

3.1 Method

The speech stimuli were English words and non-words used in the reading test 1 spoken by a male speaker of native American English. These utterances were recorded in a computer at an 11.025 kHz sampling rate. The subjects were Group A in the reading test 1. They were instructed to listen to one of these words and non-words through the headphone, and immediately repeat it three times as exactly as possible. Their utterances were recorded at an 11.025 kHz sampling rate by a solid state recorder (Marantz PMD671). The first utterances were measured in general, but the first utterance was not correctly pronounced (e. g.

(a)

(b)

dra

gra

s t r e i t

s t o r e i t o

d o r a

g u r a

Acoustics 08 Paris

653

“bra” was pronounced as /baa/), second or third one was measured. Epenthetic vowels if any and their lengths were measured by visual inspection of the spectrogram of the best utterance on a computer.

3.2 Results

The results were shown in Table 4. In most of the utterances, with few exceptions Japanese speakers did not produce epenthetic vowels in repetition test. Moreover, the length of exceptional epenthetic vowels found under the repetition test was shorter than those under the reading test 1. Figure 3 shows a spectrogram of “straight” of subject E. By comparison with a spectrogram of subject E in the reading test 1, Figure 1(b), no vowels were apparently observed in the consonant cluster /str/. Figure 4 shows the spectrograms of two subjects. In both subjects there are no vowels in consonant cluster /dm/ in non-word “edmo”. If native Japanese speakers heard non-existent vowels in consonant clusters, they would insert vowels in consonant clusters in repetition test. In the repetition test, vowel epenthesis seldom occurred, therefore it is inferred that the native Japanese speakers did not hear non-existent vowels in consonant clusters.

meandrill 76.6

drama 53.4drive 46.3dress 58.0dry 46.8

strike 52.4straight 52.0stripe 53.1

trouble 28.0truck 35.5trend 38.6try 32.0

edmo 62.3etmo 53.1dra 70.2tra 58.1gra 71.6cra 52.1bra 67.2pra 51.5

Table 3 Pooled data of the reading test 1.

Figure 3 a spectrogram of “straight” spoken by subject E in

repetition test.

Figure 4 Spectrograms of “edmo” in repetition test.

(a) subject B, (b) subject D.

A B C D E F G Hdrill 0 0 0 0 0 0 0 [u] 58

drama 0 0 0 [o] 43 0 0 0 0drive x 0 0 0 0 0 0 [o] 30dress [o] 30 0 0 0 0 0 0 [o] 37dry 0 0 0 0 0 0 0 [u] 43

strike 0 0 0 0 0 0 0 xstraight [o] 22 0 0 0 0 0 0 0stripe [o] 15 [o] 20 x 0 0 0 [o] 30 0

trouble 0 0 0 0 0 0 0 0truck 0 0 [o] 18 0 0 0 0 0trend 0 0 0 0 0 0 0 0try 0 0 0 0 0 0 0 [o] 40

edmo 0 0 0 0 0 0 0 [u] 43etmo 0 0 [o] 43 0 [e] 30 0 0 0dra 0 0 0 0 0 0 0 0tra 0 0 0 0 0 0 x 0gra 0 0 0 0 0 0 0 [u] 52cra 0 0 0 0 0 x 0 0bra x 0 0 x 0 x 0 [u] 60pra x 0 0 x 0 x x x

Table 4 Epenthetic vowel lengths in repetition test. Subjects A~H correspond to Table 1.

4 Reading test 2

4.1 Method

Reading test 2 was done after 5 or 6 months after reading test 1. Subjects were Group A in the reading test 1. Speech samples and procedures were the same as for the reading test 1.

4.2 Results

Results are shown in Table 5. In comparison with Table 1, as in the case of the reading test 1, most of the subjects

(b)

(a)

s t r e i t

e d m o

e d m o

Acoustics 08 Paris

654

inserted vowels in consonant clusters. Although subjects did not insert vowels in the repetition test, nevertheless they inserted vowels this time. Our subjects were novices at English conversation, not good at speaking English. Therefore, we might suppose that when they looked at an English word, they associated it with a corresponding loan word. Consequently, they might insert vowels in consonant clusters.

A B C D E F G H meandrill 104 76 85 82 63 [u] 19 83 138 90.1

drama 62 55 45 52 34 26 61 99 54.3drive 37 53 48 42 39 [u] 19 56 63 48.3dress 50 45 30 37 65 0 56 99 47.8dry 40 47 35 49 61 46 46 77 50.1

strike 39 42 53 40 56 0 70 60 45.0straight 50 43 50 38 61 25 68 62 49.6stripe 57 52 x 38 70 0 56 x 45.5

trouble 50 22 0 26 58 0 63 25 30.5truck [*] 30 12 0 21 22 0 26 80 23.0trend [*] 45 18 0 28 27 0 54 43 24.3try 44 37 25 33 17 15 54 43 33.5

edmo 74 58 55 57 49 66 44 [u] 95 57.6etmo 74 0 74 49 63 53 60 [u] 90 53.3dra 99 58 60 56 73 61 77 88 71.5tra 70 0 70 42 66 [*] 43 66 106 60.0gra 59 73 81 56 71 75 70 61 68.3cra 28 0 60 35 44 31 53 56 38.4bra 104 108 23 40 85 80 54 105 74.9pra 54 0 35 45 36 48 53 76 43.4 Table 5 Epenthetic vowel lengths in the reading test 2.

Subjects A~H correspond to Table 1. [*] = [o]

5 Magnetoencephalographic test

5.1 Method

The subjects were native Japanese speakers. All subjects were right-handed, and had no hearing loss. Subjects were instructed not to attend to the stimuli and to hear passively to the stimulus sequence. Stimuli were binaurally presented to subjects by inserted earphones. Stimuli, /dra/, /dora/, /bra/, and /bura/ were spoken by a male speaker of Japanese. An oddball paradigm was adopted, in which a standard stimulus was presented at a high frequency of 85% and deviant at a low frequency of 15%. Deviants never occurred in immediate succession. Stimulus onset asynchrony was one second. Four sessions with a standard-deviant pair set as /dra/ vs. /dora/, /dora/ vs. /dra/, /bra/ vs. /bura/ and /bura/ vs. /bra/, respectively, were conducted in a counter-balanced order. The recordings were performed in a magnetically shielded room using a 204-channel whole head magnetometer (Neuromag Ltd., Finland). Mismatch fields (MMF), which is an event-related neural response reflecting automatic detection of acoustic changes, were analyzed. MMF was determined from the deviant stimulus response minus the standard stimulus response subtraction waves. For each subject and condition, equivalent current dipoles (ECDs) were determined for the MMF.

5.2 Results

Figure 5 shows a magnetic field of /dora/ deviant in one subject and distinct MMF is observed. Figure 6 shows the

ECD moment in dr (/dra/ vs. /dora/ and /dora/ vs. /dra/) and br (/bra/ vs. /bura/ and /bura/ vs. /bra/) conditions. In all conditions MMFs were generated, hence it is apparent that subjects detected vowel insertion in consonant clusters.

Figure 5 Magnetic fields of deviant “dora”.

Red line: MMF, blue dashed line: deviant stimulus response, black dotted line: standard stimulus response.

05

1015202530

br dr

RHLH

Figure 6 ECD moments.

LH: left hemisphere, RH: right hemisphere.

6 Discussion

In the reading test, the native Japanese speakers inserted a vowel in the consonant clusters as required in Japanese phonotactics, whereas in the repetition test, they did not insert any vowel in the consonant clusters with few exceptions. Moreover, the exceptionally inserted vowels in the repetition test were shorter than those in the reading test 1. If they perceived a non-existent vowel in the consonant clusters, they would insert a vowel in the consonant clusters even in the repetition test. Moreover, the lengths of the inserted vowels would be the same as those in the reading test 1. The magnetoencephalographic test showed that since apparent MMFs were generated, epenthetic vowels in consonant clusters were detected by Japanese subjects. According to the above results, we interpret our results that native Japanese speakers can detect epenthetic vowels in consonant clusters and can correctly pronounce consonant clusters, but their detection sensitivity for a vowel segment in a consonant cluster may be lower than French subjects. When Japanese subjects read loan words, they tend to insert a vowel in consonant clusters following the phonological constraints of Japanese which do not allow consonant clusters. Japanese speakers need to perceive the same words whether a vowel exists in a consonant cluster or not. The sensitivity of Japanese speakers to detect vowel deletion or

100 ms

100 fT/cm

mom

ent (

nAm

)

Acoustics 08 Paris

655

insertion within a consonant cluster does not decline at least at the acoustical level, but is suppressed at the word recognition level following the phonological constraints of Japanese.

7 Conclusion

The results of the present study suggest that native Japanese speakers can detect epenthetic vowels in consonant clusters and can correctly pronounce consonant clusters. The sensitivity of Japanese speakers to detect vowel deletion or insertion within a consonant cluster does not decline at least at the acoustic perceptual level, but seems to be suppressed at the word recognition level to follow the phonological constraints of Japanese.

Acknowledgments

This study was supported by Grant-in-Aid for Scientific Research of JSPS (No. 18520327).

References

[1] L. Davidson, “Addressing phonological questions with ultrasound”, Clinical Linguistics & Phonetics, 19, 619-33 (2005)

[2] E. Dupoux, K. Kakehi, Y. Hirose, C. Pallier, J. Mehler, “Epenthetic vowels in Japanese: A perceptual illusion?”, Journal of Experimental Psychology: Human Perception & Performance 25, 1568-1578 (1999)

[3] E. Dupoux, C. Pallier, K. Kakehi, J. Mehler, “New evidence for prelexical phonological processing in word recognition”, Language & Cognitive Processes, 16, 491-505 (2001)

[4] G. Dehaene-Lambertz, E. Dupoux, A. Gout, “Electrophysiological correlates of phonological processing: A cross-linguistic study”, Journal of Cognitive Neuroscience, 12, 635-47 (2000)

[5] H. Quackenbush, M. Ooso, “Gairaigo no keisei to sono kyooiku”, The National Institute for Japanese Language, (1990)

[6] H. Kubozono, “Nihongo no onsei”, Iwanami, Tokyo, Japan, (1999)

[7] H. Kubozono, “Prosodic structure of loanwords in Japanese: Syllable structure, accent and morphology”, Journal of Phonetic Society of Japan, 6, 79-97 (2002)

[8] K. Maekawa, H. Kikuchi, “Corpus-based analysis of vowel devoicing in spontaneous Japanese: an interim report”, In Voicing in Japanese, edited by J. van de Weijer, K. Nanjo, T. Nishida, Mouton de Gruyter, Berlin, Germany, (2005)

Acoustics 08 Paris

656